Real-Time Deepfake Detection for AI-Generated Arabic Speech

Abstract

This study investigates the effectiveness of Retrieval-based Voice Conversion (RVC) in detecting AI-generated Arabic speech across diverse linguistic contexts. The primary research questions address whether the RVC model can accurately differentiate between real and synthetic speech samples in various languages and its generalization capability across different linguistic contexts. Experimental evaluations show that the model achieves accuracies of 92% using Ensemble Learning (XGBoost and LightGBM), 93% with Meta Learning, and 90% with Neural Networks. Assessments encompass scenarios with real Arabic and English speech not included in training datasets, as well as different speakers across languages. The study contributes insights into the robustness and practical application of RVC for speech authentication.